Billions of sensors used at enterprises nowadays collect physical environment information and generate large amounts of data, which, combined with the power of AI, can be used to solve many hard problems in a way we could not even imagine just some years ago. Often, such AI solutions require storage and processing capabilities close to where the data is generated — at the edge. Unlike a physical data center, edge devices have lower computing, limited storage and restricted power consumption. AI networks, in turn, become larger and more complex by encoding more information. Such an increase in size and complexity is, in turn, associated with AI models having lower throughput and larger memory requirements. With that, real-time AI inference is becoming a new great challenge for embedded data processing. In this talk, Ekaterina will observe trends and challenges in embedded AI for machine vision use cases and will share some best practices for optimizing AI inference at the edge.